search

kevaldoshi17 / Prediction-based-Video-Anomaly-Detection-

MIT License
17 stars 10 forks source link

Low FPS when running in Docker #7

Open mausLe opened 3 years ago

mausLe commented 3 years ago

I'm running the MONAD framework on Docker with the nvidia/cuda:10.2-cudnn8-devel-ubuntu18.04 and my specs are RTX2080 - CUDA 11.2

But it only achieves 2FPS (I disabled YOLO, just kept the GAN model). As I checked GPU, it didn't consume any of it.

Do you know what is the problem?

Thanks for your help.

kevaldoshi17 commented 3 years ago

Hi,

This most likely seems to be a CUDA/Nvidia driver mismatch issue. Can you please check if tensorflow is recognizing the gpu or not?

kevaldoshi17 commented 3 years ago

Also please check if you have the right version of CUDNN installed.

mausLe commented 3 years ago

Thanks for replying.

I pull the new image nvidia/cuda:10.0-cudnn7-devel-ubuntu18.04 and it finally works now.

The problem here was that I used Tensorflow 1.15.2 which is compatible with cuDNN 7 and CUDA 10.0. While my previous docker image, cuda:10.2-cudnn8 failed to run MONAD on the Tensorflow-GPU and ran it on Tensorflow-CPU instead. Which caused the FPS to be so low.

mausLe commented 3 years ago

Btw, could you mind explain to me how could I calculate the lambda lm2 value for each frame episode?

If I test in on the new data, how do I apply this lm2 value?

Thank you.

monad_lambda_calc

monad_lambda